Pith. sign in

REVIEW 4 major objections 5 minor 50 references

NER4all or Context is All You Need: Using LLMs for low-effort, high-performance NER on historical texts. A humanities informed approach

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Prompting an off-the-shelf large language model with historical context and a historian's persona outperforms specialist NER frameworks on a 1921 German travel guide, lifting F1 scores roughly 7 to 22 percent above the spaCy and flair…

desk verdict Useful empirical study with one solid surprise (zero-shot beats few-shot until 16 examples), but the headline F1 advantage over spaCy/flair is compromised by a missing development/evaluation split. read the letter →

arxiv 2502.04351 v1 pith:VM3PUQ63 submitted 2025-02-04 cs.CL cs.AI

Torsten Hiltmann , Martin Dröge , Nicole Dresselhaus , Till Grallert , Melanie Althage , Paul Bayer , Sophie Eckenstaler , Koray Mendi
show 4 more authors
Jascha Marijn Schmitz Philipp Schneider Wiebke Sczeponik Anica Skibba
This is my paper · ORCID
classification cs.CLcs.AI
keywords namedentityrecognitionhistoricaldocumentslargelanguagemodelspromptengineeringpersonamodelingzero-shotlearningfew-shotdigitalhumanities
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper aims to show that named entity recognition (NER) on historical texts can be done with a consumer-grade large language model and a carefully written prompt, without fine-tuning or training data. On 25 randomly selected pages of the 1921 Baedeker travel guide 'Berlin und Umgebung', the authors report that ChatGPT-4o with a context-rich, persona-based 'full prompt' reaches F1 scores of about 0.87-0.88, compared with 0.81 for the established framework flair and 0.66 for spaCy. The paper's key argument is that providing specific historical context and casting the model as an expert historian activates knowledge the LLM already has, turning NER into a content-oriented task rather than a purely linguistic one. The authors also report that zero-shot prompting beats one- to eight-shot prompting, with improvement appearing only at sixteen or more examples, and that English and German prompts perform comparably once the full prompt engineering is applied. A sympathetic reader would take away that the bottleneck for historical NER has shifted from expensive annotation and model training to prompt design.

What carries the argument

The load-bearing mechanism is the 'full prompt', a structured natural-language instruction set that combines source-specific historical context about the 1921 Baedeker guide and Berlin around 1920, an expert persona ('experienced linguist specializing in NER for historical texts and expert in various historical epochs and locations'), a span-tag output format in which entities are wrapped as <<TAG ... /TAG>>, NOTE rules that correct systematic errors the authors observed (Greek gods and heroes count as people; restaurants count as organizations even if person-named; locations inside quoted entities are still annotated), instruction repetition, reward/punishment incentives, and the phrase 'take a deep breath and think step by step.' The prompt is designed to activate the historical and world knowledge the LLM absorbed during training, so that annotation proceeds from content understanding rather than surface linguistic pattern matching. A fuzzy-search post-processor then locates the generated spans in the source text, allowing about one character error per five generated characters, and the evaluation uses overlap-based entity-type matching to compare spans with ground truth.

What would settle it

Freeze the finalized full prompt and apply it to a freshly annotated set of pages from the same Baedeker edition (or another historical corpus) that was never seen during prompt development, using the same fuzzy span matching; if the F1 margin over flair falls from roughly 0.06 to within the page-wise standard deviation of about 0.08, the 7-22 point advantage would be contradicted as a generalizable claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that an out-of-the-box commercial LLM, gpt4o-2024-08-06, significantly outperforms two leading specialist NER frameworks on historical documents when the task is reframed with historical context and persona modeling. In the paper's evaluation on 25 pages of the 1921 Baedeker guide, the full prompt reaches recall 0.84-0.85, precision 0.91, and F1 0.87 (German) or 0.88 (English), against flair's 0.76/0.89/0.81 and spaCy's 0.71/0.62/0.66. Generic historical context improves recall over a no-context prompt by about six percentage points in the German zero-shot condition, source-specific context adds further gains, and the authors attribute the remaining lift to prompt-engineering techniques such as instruction repetition, system-prompt separation, and the instruction to think step by step. The paper also claims that examples are not the main driver: zero-shot outperforms one-, two-, four-, and eight-shot prompts, and only the sixteen- and thirty-two-shot conditions match or exceed the zero-shot F1. The authors interpret all of this as evidence that the model's latent historical knowledge, mobilized by context, is what carries performance.

Load-bearing premise

The conclusion stands on the assumption that the final prompt—its context wording, NOTE rules, and engineering tricks—was chosen without using the same 25 pages that were later scored; if those pages shaped the prompt, the reported gains over flair and spaCy would partly be test-set fitting and would not transfer to new texts.

Editorial extensions

If this is right

  • Historians who can write a natural-language prompt can obtain NER results on historical texts that match or exceed specialist NLP tools, removing the need for scripting skills, model training, or large annotated datasets.
  • For annotation budgets, the paper's results imply that context writing and error-note design are more cost-effective than hand-labeling example spans, since zero-shot beats one-to-eight-shot and only 16+ examples help.
  • Because English and German prompting converge once the full prompt is used, the approach is accessible to non-English-speaking researchers without an accuracy penalty.
  • Preliminary tests on 16th-18th century documents of various genres suggest the method transfers beyond the Baedeker corpus, though the paper flags that further testing is needed.
  • The finding that small example sets can hurt recall implies that few-shot prompting practices from modern-text NER should not be imported uncritically into historical-text workflows.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the context-activation mechanism suggests a general recipe for other low-resource text domains—write a paragraph that names the domain's interpretive conventions rather than collecting gold labels; a cross-corpus frozen-prompt evaluation would test whether the recipe holds.
  • Editorial inference: because the NOTE rules were adapted to errors the authors noticed, the method is partly corpus-specific; a stronger test of portability would write the notes on an independent development set and score on a truly unseen set.
  • Editorial inference: the 16-shot threshold raises a testable hypothesis that the crossover point depends on example length, class balance, and the typological distance between examples and target text, which future work could map systematically.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a prompting strategy for zero- to few-shot NER on historical German texts using ChatGPT-4o, arguing that domain-specific context and persona modeling turn NER from a purely linguistic task into a humanities-informed task. The authors evaluate their approach on 25 pages of a 1921 Baedeker travel guide with manually annotated ground truth, comparing it to off-the-shelf flair and spaCy baselines. They report F1 scores of 0.87–0.88 for their full prompt versus 0.81 for flair and 0.66 for spaCy, and they present an ablation study of context levels and prompt-engineering components. They also report that zero-shot outperforms few-shot until 16–32 examples are provided and that English and German prompts perform similarly once prompt engineering is used.

Significance. If the central result holds, the paper makes a useful empirical contribution to historical NER: a low-effort, prompt-only approach that outperforms off-the-shelf frameworks on a realistic historical corpus, with an ablation isolating the contributions of context, persona, and prompt-engineering tricks. The most interesting finding, that few-shot examples do not help below 16 shots, is potentially valuable for practitioners who cannot afford expert-annotated examples. The authors also provide data and code on OSF, which supports reproducibility. However, the main quantitative claim is currently vulnerable because the manuscript does not document that prompt development was separated from the evaluation pages, and no significance testing or strict-match sensitivity analysis is provided.

major comments (4)
  1. [Results / Table 2] The manuscript does not document a separate development set. In the Method section the authors state that they 'chose the best combination that worked during experimentation without a comprehensive study', and in 'Instructing the LLM' they report that the NOTE rules were added to address 'systematic errors we noticed' and that 'all notes ought to be adapted to the specific, observed errors of the used LLM'. The evaluation uses 25 randomly selected pages from the same 55-page annotated set described in 'Corpus and data set', and the paper never states that prompt development was confined to the remaining 30 pages. If the context wording and NOTE rules were selected after inspecting outputs on the 25 evaluation pages, the reported 0.87 vs. 0.81 F1 advantage over flair is partly an artifact of test-set fitting and the headline claim would not be established. The revision must either document a clean development/evaluation split or rerun the comparison with a properly held-out division.
  2. [Results / Table 2] The claim that zero-shot outperforms few-shot until 16 examples is not supported by significance testing. In Table 2, the 0-shot German F1 of 0.87 ± 0.08 differs from the 1-shot German F1 of 0.86 ± 0.08 by far less than the reported standard deviations, and the 16-shot F1 of 0.88 ± 0.07 is within one standard deviation of the 0-shot value. The same applies to the claim that English and German full prompts perform 'insignificantly' differently. Without paired statistical tests or confidence intervals, the ordering of these conditions may simply reflect noise. The revision should add appropriate tests or soften the conclusions to what the data can actually support.
  3. [Matching spans through post-processing the results] The evaluation uses the 'most lenient' NER matching criterion (ent_type with overlap) and a fuzzy-search tolerance of one Levenshtein error per five characters before matching. This can inflate absolute F1 scores and could differentially favor the LLM if its output spans are longer or shifted relative to the baselines. The paper should report a strict or partial matching sensitivity analysis, and it should state explicitly whether the flair and spaCy predictions were subjected to exactly the same fuzzy-search and overlap-matching pipeline as the LLM outputs. Without this, the absolute F1 values and the size of the reported gap over the baselines are hard to interpret.
  4. [Transferability and future work] The transferability paragraph asserts that 'preliminary tests on documents from the 16th to the 18th centuries ... suggest that our methodology can be applied to texts from different historical periods and backgrounds with clear improvements over the respective baselines', but no numbers, corpora, or evaluation details are given. This is an unsupported empirical claim in an otherwise quantitative paper. The authors should either remove the quantitative wording ('clear improvements') or report the actual results, even if preliminary.
minor comments (5)
  1. [Selection of LLMs / Method] The model identifier is inconsistent: the Method section names 'gpt4o-2024-08-16', while the 'Selection of LLMs' section names 'gpt4o-2024-08-06'. Because the exact model snapshot affects reproducibility, this should be corrected and used consistently.
  2. [Abstract] The abstract contains typos and grammatical errors that should be fixed, including 'do to' (should be 'due to'), 'democratises' (consistent UK spelling is fine but the surrounding text is mixed), and the phrase 'seven to twentytwo percent' should be clarified as percentage points or relative percent because the tables report F1 differences of 0.06–0.22 on a 0–1 scale.
  3. [Results / Prompting language] The text states that 'this expectation was confirmed and yielded a 5% improvement in recall' for English prompts, but the reported tables do not show a 5% recall gap (Table 2 shows 0-shot English recall 0.85 versus German 0.84). The manuscript should clarify which comparison this 5% figure refers to, or correct the number.
  4. [Evaluation of performance of prompt engineering] The discussion of the spaCy baseline is confusing regarding the reported F1: the text cites a self-reported F1 of 85 for de_core_news_lg, while footnote 3 mentions a 91.6 F1 for a RoBERTa-based model. Please clarify which number corresponds to which model and why only the lower value is used in the comparison.
  5. [Including examples] The notation '2ˆ1' and '2ˆ5' in the 'Including examples' section appears to intend '2^1' and '2^5' but uses a circumflex with a Unicode superscript; this should be typeset consistently, and the sentence 'This is drastically lower than the whole page' should be reworded for clarity.

Circularity Check

1 steps flagged · score 4.0 of 10

Prompt components were explicitly selected during experimentation and adapted to 'observed errors,' but the paper never documents a held-out development set; the reported F1 advantage over flair is partly a fitted value, though external baselines and human ground truth provide independent content.

  1. fitted input called prediction [Method: 'Method' and 'Instructing the LLM' (arXiv pp. 8-11); evaluation corpus of 25 pages defined in 'Corpus and data set' (p. 8)]
    "Because there are a lot of combinations for different prompting techniques, we chose the best combination that worked during experimentation without a comprehensive study. ... The 'NOTE' parts address systematic errors we noticed ... All notes ought to be adapted to the specific, observed errors of the used LLM to improve performance."

    The paper's headline result (full prompt F1 0.87-0.88 vs. flair 0.81 and spaCy 0.66) is produced by a prompt whose components were chosen during 'experimentation' and whose NOTE rules were explicitly added to correct 'systematic errors we noticed.' No separate development set is described anywhere; the only evaluation corpus described is the 25 pages used for the reported scores. On the paper's own account, the NOTE rules and the 'best combination' are parameters fitted to observed LLM errors, and the same observed errors are then used to compute the reported F1. The paper thus presents a fitted prompt's performance as a measured, generalizable prediction.

full rationale

There is no self-citation chain, no imported uniqueness theorem, and no definitional equation equating the output to the input. The comparison against human annotations and against two external NER frameworks (flair, spaCy) is genuine and gives the paper substantial independent content: even the less-engineered context prompts outperform spaCy and reach or exceed flair. The circularity concern is narrower but real: the prompt, especially the NOTE rules, was explicitly tuned to 'systematic errors we noticed,' and the paper never documents that this tuning was confined to a development set disjoint from the 25 evaluation pages. On the text as written, the reported 7-point F1 gain over flair comes from a prompt fitted to the same corpus on which it is evaluated, making that specific quantitative claim a fitted value rather than a prediction. This supports a moderate circularity score, not a high one, because the overall method is still anchored in external benchmarks and human ground truth.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The empirical claims rest on the correctness of the manual annotation, on the representativeness of a single 1921 travel guide and one ChatGPT-4o snapshot, and on the assumption that prompt selection was independent of the evaluation pages. No inter-annotator agreement or significance tests are reported.

free parameters (5)
  • Full prompt component set = structure + instruction repetition + reward/punishment + system/user split + 'take a deep breath'
    Hand-chosen after experimentation; no separate development set documented, so selection may be fitted to the evaluation corpus.
  • Specific context paragraph = 1921 Baedeker Berlin travel guide description
    Authored from the authors' domain knowledge of the corpus; central variable in the ablation.
  • NOTE correction rules = e.g., Greek heroes as PER, restaurants as ORG, LOC inside quotes
    Added in response to observed systematic errors of the model on this corpus; evidence of tuning to the data.
  • Fuzzy matching tolerance = len(span)//5 errors
    Chosen so all generated spans could be located; the paper reports no locating errors.
  • Entity matching criterion = nervaluate ent_type overlap
    Most lenient matching; may inflate F1 by rewarding partial spans for all systems.
assumptions (5)
  • domain assumption The manual ground-truth annotations are correct and consistent.
    Double annotation with third-opinion reconciliation, but no inter-annotator agreement reported.
  • domain assumption The 25 evaluation pages are representative of historical documents generally.
    Single corpus of a 1921 Baedeker travel guide, only 55 pages annotated, 25 evaluated; transferability claims are preliminary.
  • domain assumption ChatGPT-4o is representative of state-of-the-art LLMs.
    Only one proprietary model snapshot is used; the paper text inconsistently names 2024-08-06 and 2024-08-16.
  • domain assumption Prompt selection was independent of the evaluation pages.
    Not stated; the Method describes choosing the best prompt during experimentation and adding NOTE rules for observed errors.
  • domain assumption Lenient overlap matching is a valid basis for comparing NER systems.
    ent_type overlap may credit partial span matches; consistent across systems but inflates absolute scores.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NER4all or Context is All You Need: Using LLMs for low-effort, high-performance NER on historical texts. A humanities informed approach." pith.science (2026). https://pith.science/paper/VM3PUQ63

@misc{pith2026250204351,
  author       = {Pith},
  title        = {Pith review of: NER4all or Context is All You Need: Using LLMs for low-effort, high-performance NER on historical texts. A humanities informed approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VM3PUQ63}},
  note         = {Machine review of arXiv:2502.04351}
}
read the original abstract

Named entity recognition (NER) is a core task for historical research in automatically establishing all references to people, places, events and the like. Yet, do to the high linguistic and genre diversity of sources, only limited canonisation of spellings, the level of required historical domain knowledge, and the scarcity of annotated training data, established approaches to natural language processing (NLP) have been both extremely expensive and yielded only unsatisfactory results in terms of recall and precision. Our paper introduces a new approach. We demonstrate how readily-available, state-of-the-art LLMs significantly outperform two leading NLP frameworks, spaCy and flair, for NER in historical documents by seven to twentytwo percent higher F1-Scores. Our ablation study shows how providing historical context to the task and a bit of persona modelling that turns focus away from a purely linguistic approach are core to a successful prompting strategy. We also demonstrate that, contrary to our expectations, providing increasing numbers of examples in few-shot approaches does not improve recall or precision below a threshold of 16-shot. In consequence, our approach democratises access to NER for all historians by removing the barrier of scripting languages and computational skills required for established NLP tools and instead leveraging natural language prompts and consumer-grade tools and frontends.

Figures

Figures reproduced from arXiv: 2502.04351 by the authors.

Figure 1
Figure 1. Main instructions in the German prompt Overview of the prompt structure The main instructions (2) are as follows (fig.1, note that the screenshot shows the German prompt): 1 Follow these steps: 2 10 [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 32 canonical work pages

  1. [1]

    Akbik, Alan, Tanja Bergmann, Duncan Blythe, Kashif Rasul, Stefan Schweter, and Roland Vollgraf. 2019. `` FLAIR : An Easy-to-Use Framework for State-of-the-Art NLP .'' In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics ( Demonstrations ) , 54--59. Minneapolis: Association for Computational L...

    doi:10.18653/v1/n19-4010 2019

  2. [2]

    Baedeker, Karl, and Bruno Graupe. 1921. Berlin Und Umgebung : Handbuch Für Reisende . 19th ed. Leipzig: Karl Baedeker. https://nbn-resolving.org/urn:nbn:de:kobv:11-717582

    work page 1921

  3. [3]

    (2019) 2020

    Batista, David, and Matthew Antony Upson. (2019) 2020. ``Nervaluate.'' https://github.com/mantisnlp/nervaluate

    work page 2019

  4. [4]

    Brown, Tom B., Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, et al. 2020. ``Language Models Are Few-Shot Learners .'' July 22, 2020. https://doi.org/10.48550/arXiv.2005.14165

    arXiv doi:10.48550/arxiv.2005.14165 2020

  5. [5]

    Bsharat, Sondos Mahmoud, Aidar Myrzakhan, and Zhiqiang Shen. 2024. ``Principled Instructions Are All You Need for Questioning LLaMA-1 /2, GPT-3 .5/4.'' January 18, 2024. http://arxiv.org/abs/2312.16171

    arXiv 2024

  6. [6]

    Chen, Jiawei, Yaojie Lu, Hongyu Lin, Jie Lou, Wei Jia, Dai Dai, Hua Wu, Boxi Cao, Xianpei Han, and Le Sun. 2023. ``Learning In-context Learning for Named Entity Recognition .'' May 26, 2023. https://doi.org/10.48550/arXiv.2305.11038

    arXiv doi:10.48550/arxiv.2305.11038 2023

  7. [7]

    Lasko, Qiaozhu Mei, Joshua C

    Chen, Yukun, Thomas A. Lasko, Qiaozhu Mei, Joshua C. Denny, and Hua Xu. 2015. ``A Study of Active Learning Methods for Named Entity Recognition in Clinical Text.'' Journal of Biomedical Informatics 58 (December): 11--18. https://doi.org/10.1016/j.jbi.2015.09.010

    work page doi:10.1016/j.jbi.2015.09.010 2015

  8. [8]

    De Toni, Francesco, Christopher Akiki, Javier de la Rosa, Clémentine Fourrier, Enrique Manjavacas, Stefan Schweter, and Daniel van Strien. 2022. ``Entities, Dates , and Languages : Zero-Shot on Historical Texts with T0 .'' http://arxiv.org/abs/2204.05211

    work page Pith review arXiv 2022

Show all 50 references
  1. [9]

    Explosion

    ``De\_core\_news\_lg.'' 2023. Explosion. https://github.com/explosion/spacy-models/releases/tag/de_core_news_lg-3.7.0

    2023

  2. [10]

    Ehrmann, Maud. 2008. ``Les Entitées Nommées, de la linguistique au TAL : Statut théorique et méthodes de désambiguïsation.'' PhD thesis, Paris Diderot University. https://hal.science/tel-01639190

    2008

  3. [11]

    Ehrmann, Maud, Ahmed Hamdi, Elvys Linhares Pontes, Matteo Romanello, and Antoine Doucet. 2023. ``Named Entity Recognition and Classification in Historical Documents : A Survey .'' ACM Computing Surveys 56 (2): 27:1--47. https://doi.org/10.1145/3604931

    2023 doi

  4. [12]

    Ehrmann, Maud, Camille Watter, Matteo Romanello, Simon Clematide, and Flückiger. 2020. ``Impresso Named Entity Annotation Guidelines ,'' January. https://doi.org/10.5281/zenodo.3604227

    2020 doi

  5. [13]

    Eijnatten, Jorsi van. 2024. ``Dutch Intellectual Culture Between 1962 and 1995, or, Using Classical Algorithms and LLMs to Efficiently Extract Data with Imperfect OCR .'' In. https://2024.dhbenelux.org/wp-content/uploads/2024/05/DHB24_paper_van_Eijnatten_Dutch-intellectual-cul...

    2024

  6. [14]

    Federiakin, Denis, Dimitri Molerov, Olga Zlatkin-Troitschanskaia, and Andreas Maur. 2024. ``Prompt Engineering as a New 21st Century Skill.'' Frontiers in Education 9 (November): 1366434. https://doi.org/10.3389/feduc.2024.1366434

    2024

  7. [15]

    Giovanelli, Riccardo, and Arianna Traviglia. 2024. `` AIKoGAM : An AI-driven Knowledge Graph of the Antiquities Market : Toward Automatised Methods to Identify Illicit Trafficking Networks .'' Journal of Computer Applications in Archaeology 7 (January): 92--114. https://doi.org/gtr3s6

    2024

  8. [16]

    Moreno, and Antoine Doucet

    González-Gallardo, Carlos-Emiliano, Emanuela Boros, Nancy Girdhar, Ahmed Hamdi, Jose G. Moreno, and Antoine Doucet. 2023. ``Yes but.. Can ChatGPT Identify Entities in Historical Documents ?'' 2023 ACM/IEEE Joint Conference on Digital Libraries (JCDL), June, 184--89. https://do...

    2023

  9. [17]

    Graham, Shawn, Donna Yates, and Ahmed El-Roby. 2023. ``Investigating Antiquities Trafficking with Generative Pre-Trained Transformer ( GPT )-3 Enabled Knowledge Graphs: A Case Study.'' Open Research Europe 3 (100). https://doi.org/gtbsdp

    2023

  10. [18]

    Holla, Kiran Voderhobli, Chaithanya Kumar, and Aryan Singh. 2024. ``Large Language Models Aren't All That You Need.'' https://doi.org/10.48550/ARXIV.2401.00698

    2024 doi arXiv

  11. [19]

    Honnibal, Matthew, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. 2020. `` spaCy : Industrial-strength Natural Language Processing in Python.'' https://doi.org/10.5281/zenodo.1212303

    2020 doi

  12. [20]

    Karjus, Andres. 2023. ``Machine-Assisted Mixed Methods: Augmenting Humanities and Social Sciences with Artificial Intelligence.'' September 24, 2023. https://doi.org/10.48550/arXiv.2309.14379

    2023 doi arXiv

  13. [21]

    Keraghel, Imed, Stanislas Morbieu, and Mohamed Nadif. 2024. ``A Survey on Recent Advances in Named Entity Recognition.'' January 19, 2024. https://doi.org/10.48550/arXiv.2401.10825

    2024 doi arXiv

  14. [22]

    Klie, Jan-Christoph, Michael Bugert, Beto Boullosa, Richard Eckart de Castilho, and Iryna Gurevych. 2018. ``The INCEpTION Platform: Machine-assisted and Knowledge-Oriented Interactive Annotation.'' In Proceedings of the 27th International Conference on Computational Linguistic...

    2018

  15. [23]

    Levenshtein, V. I. 1966. ``Binary Codes Capable of Correcting Deletions , Insertions and Reversals .'' Soviet Physics Doklady 10 (February): 707. https://ui.adsabs.harvard.edu/abs/1966SPhD...10..707L

    1966

  16. [24]

    Levy, Itay, Ben Bogin, and Jonathan Berant. 2023. ``Diverse Demonstrations Improve In-context Compositional Generalization .'' In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics ( Volume 1: Long Papers ) , 1401--22. Toronto, Canada: Asso...

    2023 doi

  17. [25]

    Liu, Jiachang, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. 2021. ``What Makes Good In-Context Examples for GPT-3 ?'' January 17, 2021. http://arxiv.org/abs/2101.06804

    2021 arXiv

  18. [26]

    Nguyen, Tai, and Eric Wong. 2023. ``In-Context Example Selection with Influences .'' June 5, 2023. https://doi.org/10.48550/arXiv.2302.11042

    2023 doi arXiv

  19. [27]

    OpenAI. 2022. `` ChatGPT [ Large Language Model ] .'' https://openai.com/index/chatgpt/

    2022

  20. [28]

    ---------. 2024. `` ChatGPT-4o .'' https://openai.com/index/hello-gpt-4o/

    2024

  21. [29]

    OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, et al. 2024. `` GPT-4 Technical Report .'' March 4, 2024. http://arxiv.org/abs/2303.08774

    2024 arXiv

  22. [30]

    Pakhale, Kalyani. 2023. ``Comprehensive Overview of Named Entity Recognition : Models , Domain-Specific Applications and Challenges .'' September 25, 2023. https://doi.org/10.48550/arXiv.2309.14084

    2023 doi arXiv

  23. [31]

    Reul, Christian, Dennis Christ, Alexander Hartelt, Nico Balbach, Maximilian Wehner, Uwe Springmann, Christoph Wick, Christine Grundig, Andreas Büttner, and Frank Puppe. 2019. `` OCR4all : An Open-Source Tool Providing a ( Semi- ) Automatic OCR Workflow for Historical Printings...

    2019 doi

  24. [32]

    Reul, Christian, and Christoph Wick. 2021. ``Calamari- OCR /Calamari\_models: Pretrained Mixed Models to Be Used with Calamari .'' https://github.com/Calamari-OCR/calamari_models

    2021

  25. [33]

    Rosenzweig, Roy. 2003. ``Scarcity or Abundance ? Preserving the Past in a Digital Era .'' The American Historical Review 108 (3): 735--62. https://doi.org/10.1086/ahr/108.3.735

    2003 doi

  26. [34]

    Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, et al

    Sanh, Victor, Albert Webson, Colin Raffel, Stephen H. Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, et al. 2022. ``Multitask Prompted Training Enables Zero-Shot Task Generalization .'' March 17, 2022. https://doi.org/10.48550/arXiv.2110.08207

    2022 doi arXiv

  27. [35]

    Olival, Fátima Farrica, and Renata Vieira

    Santos, Joaquim, Helena Freire Cameron, F. Olival, Fátima Farrica, and Renata Vieira. 2024. ``Named Entity Recognition Specialised for Portuguese 18th-Century History Research.'' In. https://www.semanticscholar.org/paper/Named-entity-recognition-specialised-for-Portuguese-Sant...

    2024

  28. [36]

    Sarker, Shraboni, Ahmad Tamim Hamad, Hulayyil Alshammari, Viviana Grieco, and Praveen Rao. 2024. ``Seventeenth- Century Spanish American Notary Records for Fine-Tuning Spanish Large Language Models .'' June 9, 2024. https://doi.org/10.48550/arXiv.2406.05812

    2024 doi arXiv

  29. [37]

    Schweter, Stefan, and Alan Akbik. 2021. `` FLERT : Document-Level Features for Named Entity Recognition .'' May 14, 2021. https://doi.org/10.48550/arXiv.2011.06993

    2021 doi arXiv

  30. [38]

    Segura-Bedmar, Isabel, Paloma Mart \' nez, and Mar \' a Herrero-Zazo. 2013. `` SemEval-2013 Task 9: Extraction of Drug-Drug Interactions from Biomedical Texts ( DDIExtraction 2013).'' In Second Joint Conference on Lexical and Computational Semantics (* SEM ), Volume 2: Proceed...

    2013

  31. [39]

    Stammbach, Dominik, Maria Antoniak, and Elliott Ash. 2022. ``Heroes, Villains , and Victims , and GPT-3 : Automated Extraction of Character Roles Without Training Data .'' In Proceedings of the 4th Workshop of Narrative Understanding ( WNU2022 ) , 47--56. Seattle, United State...

    2022 doi

  32. [40]

    flairNLP/flair

    ``State-of-the- Art Models .'' (2018) 2025. flairNLP/flair. January 28, 2025. https://github.com/flairNLP/flair

    2018

  33. [41]

    Tang, Xuemei, Zekun Deng, Qi Su, Hao Yang, and Jun Wang. 2024. `` CHisIEC : An Information Extraction Corpus for Ancient Chinese History .'' April 20, 2024. https://doi.org/10.48550/arXiv.2403.15088

    2024 doi arXiv

  34. [42]

    Hasimoto

    Taori, Rohan, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Li Xuechen, Carlos Guestrin, Percy Liang, and Tatsunori B. Hasimoto. 2023. ``Alpaca: A Strong , Replicable Instruction-Following Model .'' https://crfm.stanford.edu/2023/03/13/alpaca.html

    2023

  35. [43]

    Thomas, Alan, Robert Gaizauskas, and Haiping Lu. 2024. ``Leveraging LLMs for Post-OCR Correction of Historical Newspapers .'' In Proceedings of the Third Workshop on Language Technologies for Historical and Ancient Languages ( LT4HALA ) @ LREC-COLING-2024 , edited by Rachele S...

    2024

  36. [44]

    THUDM. 2023. `` ChatGLM2-6B .'' THUKEG. https://github.com/THUDM/ChatGLM2-6B/blob/main/README_EN.md

    2023

  37. [45]

    Touvron, Hugo, Louis Martin, and Kevin Stone. 2023. ``Llama 2: Open Foundation and Fine-Tuned Chat Models .''

    2023

  38. [46]

    Wang, Shuhe, Xiaofei Sun, Xiaoya Li, Rongbin Ouyang, Fei Wu, Tianwei Zhang, Jiwei Li, and Guoyin Wang. 2023. `` GPT-NER : Named Entity Recognition via Large Language Models .'' October 7, 2023. https://doi.org/10.48550/arXiv.2304.10428

    2023 doi arXiv

  39. [47]

    Wick, Christoph, Christian Reul, and Frank Puppe. 2020. ``Calamari – A High-Performance Tensorflow-based Deep Learning Package for Optical Character Recognition .'' Digital Humanities Quarterly 14 (2). https://arxiv.org/abs/1807.02004

    2020 arXiv

  40. [48]

    Winner, Langdon. 1993. ``Upon Opening the Black Box and Finding It Empty: Social Constructivism and the Philosophy of Technology.'' Science, Technology, & Human Values 18 (3, 3): 362--78. https://doi.org/10.1177/016224399301800306

    1993 doi

  41. [49]

    Yan, Faren, Peng Yu, and Xin Chen. 2024. `` LTNER : Large Language Model Tagging for Named Entity Recognition with Contextualized Entity Marking .'' April 8, 2024. https://doi.org/10.48550/arXiv.2404.05624

    2024 doi arXiv

  42. [50]

    M&v ;| p F&. Czq ]f Dۃie.7,8

    Zixiao, Zhu, Feng Zijian, Zhou Hanzhang, Qian Junlang, and Mao Kezhi. 2024. `` MICL : Improving In-Context Learning Through Multiple-Label Words in Demonstration .'' June 16, 2024. http://arxiv.org/abs/2406.10908. CSLReferences document prompt_english.png0000664000000000000000...

    2024 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.